CN100403276C - Memory access method - Google Patents

Abstract

A memory access method is proposed to reduce or eliminate page-change operation during data reading by repeatedly storing partial data of memory page boundaries. The method comprises the following steps: determining a maximum height and a maximum width of a prediction block; a storage step, repeatedly storing partial data of an adjacent previous storage page at least one boundary of at least one storage page; and a reading step, selecting a storage page which does not generate page replacement on a repeated storage boundary according to the position of the block to be read, and reading the block. Therefore, the invention eliminates the page-changing action when reading the block data by using the repeated storage mode, thereby improving the memory bandwidth and accelerating the reading speed.

Description

memory access method

technical field

The present invention relates to a memory access method, in particular to a memory access method of a video decoding system that utilizes repeated storage of data to reduce page change action during data reading.

Background technique

In many modern video decoding systems, such as MPEG-I, MPEG-II and H261, etc., compression technology is commonly used for predictive or inter-frame coding actions, in which motion compensation is based on blocks (Block), and each Predicted Blocks are combined with Motion Vectors. The operation of motion compensation is mainly to read the predicted block according to the data of the motion vector on the reference image. In general, the memory for predicting reference pictures is quite large and is stored in dynamic random access memory (DRAM). Dynamic random access memory is composed of many storage banks, and the storage bank is composed of many storage lines. Before accessing data from a storage line, it must go through some additional time periods, such as precharging and activating a storage line, etc. . So if the current access line is different from the previous one, some additional time periods will be generated, such as page change actions. The paging action will greatly reduce the bandwidth (Bandwidth) for reading prediction blocks from the DRAM.

In order to reduce paging action, the prediction block must be distributed on as few memory pages as possible, and the prediction block must be read page by page. FIG. 1 is a schematic diagram of storing reference image data in blocks. As shown in FIG. 1 , the size of the reference image is 720*576 pixels, and each macro-block is 16*16 pixels. If the memory page size of the DRAM is 1024 bytes, each memory page can store four macroblocks. The storage method in FIG. 1 stores 4 vertical macroblocks (thinner solid line) for each storage page (thicker solid line). When the dotted predicted block A is read, the data areas A1 , A2 , A3 , and A4 of the predicted block A are distributed in four different memory pages, so three page change actions are required.

Although reading predictive blocks in a page-by-page order can largely eliminate page changes, it is still a major bandwidth constraint on DRAM for high-bandwidth applications, such as HDTV video decoding systems. bottleneck. Especially in high-frequency applications with real-time decoding limitations, due to the need for a large number of fast memory reads during the motion compensation process, too many page-changing actions may not meet the real-time decoding requirements.

Contents of the invention

In view of the above problems, the object of the present invention is to provide a memory access method, which forms a simulated memory page by repeatedly storing data at the boundary of the memory page, thereby reducing or completely avoiding the paging action.

In order to achieve the above object, the memory access method of the present invention reduces or eliminates the paging action when reading data by repeatedly storing part of the data at the memory page boundary. The method comprises the following steps: determining the maximum height and maximum width of a prediction block; storing step, repeatedly storing part of data of the adjacent previous storage page at least one boundary of at least one storage page; and reading step, according to the determined For the location of the block to be read, select a storage page that does not cause a page break on the boundary of repeated storage to read the block.

Therefore, the present invention utilizes repeated storage to eliminate the paging action when reading the block data, thereby increasing memory bandwidth and speeding up the reading speed.

Description of drawings

FIG. 1 is a schematic diagram of storing reference image data in blocks.

Figure 2 shows the principle of merging two memory pages into a simulated memory page, where (A) is the location map of block A in memory page M, and (B) is the position map of block A shifted to the right by one pixel and located in memory The position map of page M and page N, (C) is a schematic diagram of the boundary position data of storage page M that will store page N repeatedly, and (D) is a simulated storage page M_N.

3 is an example of removing all horizontal storage page boundaries according to the present invention, wherein (A) is a schematic diagram of overlapping storage of each storage page, and (B) is a simulated storage page.

4 is an example of removing all horizontal and vertical storage page boundaries according to the present invention, wherein (A) is a schematic diagram of overlapping storage of each storage page, and (B) is a simulated storage page.

Explanation of reference numbers

MB Macroblock

A prediction block

M, N memory pages

SP Emulated Memory Pages

Detailed ways

Embodiments of the memory access method of the present invention will be described in detail below with reference to the accompanying drawings.

Figure 2 shows the principle of merging two memory pages into a simulated memory page, where (A) is the location map of block A in memory page M, and (B) is the position map of block A shifted to the right by one pixel and located in memory The position map of page M and page N, (C) is a schematic diagram of the boundary position data of storage page M that will store page N repeatedly, and (D) is a simulated storage page M_N.

In FIG. 2(A), memory page M and memory page N are two horizontally adjacent memory pages located in the reference image. Since the prediction block A is completely contained by the storage page M, when the system reads the prediction block A, there will be no page change action. However, as shown in FIG. 2(B), when the prediction block A is moved to the right by one pixel, a row of data of the prediction block A is stored in the storage page N. Therefore, when the system reads the predicted block A, a page change action from page M to page N occurs. Therefore, in order to avoid the generation of the page change action, the memory access method of the present invention repeatedly stores the partial data of the storage page M at the boundary of the storage page N, that is, the data stored in the storage page N is different from the data stored in the storage page M. The stored data overlap, as shown in the hatched area in Figure 2(C). The width of the overlapping area is the width of the prediction block A minus 1.

Using this storage page overlapping technology, when the system calculates that the prediction block A will cross the vertical boundary of the storage page M, it will read the data of the prediction block A from the storage page N, and because the prediction block A It is completely contained by memory page N, so no page change is required. Therefore, when looking at the memory from the perspective of the system, the memory pages M and N are regarded as a simulated memory page M_N, as shown in FIG. 2(D). Therefore, as long as the predicted block A is located in the simulated memory page M_N, when the system reads the data of the predicted block A, no vertical boundary page change will occur. With this technique of simulating a memory page, the size of a memory page can be enlarged by duplicating data stored between adjacent memory pages. Of course, while Figure 2 is an example of removing vertical memory page boundaries, the same technique can be used to remove horizontal memory page boundaries.

In addition, the size of the region for repeated storage depends on the size of the predicted block A. If the maximum width of the prediction block is W _max and the maximum height is H _max , the unit is pixel. For example, for reference pictures stored in MPEG I or MPEG II, possible prediction block sizes with half-pixel compensation include: 16×16, 17×17, 16×17, and 17×16. Therefore, both the maximum width and the maximum height are 17, that is, W _max =17, and H _max =17.

Secondly, assuming that the capacity of the storage page is S _page , and the shape of the storage macroblock of the storage page is a rectangle, the pixels defining its width and height are respectively W _page and H _page . When removing the horizontal storage page boundary, then as shown in Figure 2 (C), the row number H _dup of the data stored repeatedly between two horizontally adjacent storage pages is (H _max -1), so H must be satisfied _page ≥ H _max . However, if the boundary of vertical storage pages is to be removed, the number of rows W _dup to be repeatedly stored between two vertically adjacent storage pages is (W _max -1), so W _page ≥ W _max must be satisfied. Therefore, the memory pages can be merged into larger simulated memory pages according to the size of the memory, thereby reducing the paging action when reading the prediction block.

3 is an example of removing all horizontal storage page boundaries according to the present invention, wherein (A) is a schematic diagram of overlapping storage of each storage page, and (B) is a simulated storage page. In the MPEG I or MPEG II specification, the size of a macroblock is 16*16 pixels, and the maximum height H _max of a prediction block is 17. If the capacity of the memory page can store 4 macroblocks, the height of the memory page H _page is 64, so the condition of H _page ≥ H _max can be satisfied, and the horizontal memory page boundaries between vertically adjacent memory pages can be removed. And the number of rows H _dup to be repeatedly stored between vertically adjacent storage pages is 16, that is, H _max -1, which is exactly the height of one macroblock. Therefore, when storing the reference image in the memory, except for the memory page of the first row, which does not need to store data repeatedly, all other memory pages repeat the bottom macroblock data of the previous memory page, as shown by the oblique line in Figure 3 area. Taking the size of the reference image as 720*576 pixels as an example, when the storage method of the present invention is not used, the number of rows of the storage page is 9 rows. After using the storage method of the present invention, since the storage page of each row must repeatedly store the data of one macroblock, the number of rows of the storage page is increased by about 1/3, that is, the total number of rows is increased to 12. Although the capacity of the memory needs to be increased, the horizontal boundary can be eliminated, as shown in FIG. 3(B), so as to improve the speed when reading the prediction block. As shown in FIG. 3(B), the horizontal boundary of each storage page has been eliminated, so the height of the storage pages SPO~SP44 defined by the system is extended. In other words, the system trades memory bandwidth for increased memory space. Secondly, since the memory page width W _page in FIG. 3(A) is smaller than the maximum prediction block width W _max , the vertical memory page boundary cannot be removed.

4 is an example of removing all horizontal and vertical storage page boundaries according to the present invention, wherein (A) is a schematic diagram of overlapping storage of each storage page, and (B) is a simulated storage page. In the MPEG I or MPEG II specification, the size of a macroblock is 16*16 pixels, and the maximum height H _max of a prediction block is 17. If the capacity of the storage page can store 16 macroblocks, the height of the storage page H _page and W _page are both 64, so the conditions of H _page ≥ H _max and W _page ≥ W _max can be satisfied, and the vertically adjacent storage pages can be removed horizontal memory page boundaries between them, and remove vertical memory page boundaries between horizontally adjacent memory pages. And the number of rows H _dup to be repeatedly stored between vertically and horizontally adjacent storage pages is 16, that is, H _max -1, which is exactly the height and width of a macroblock. Therefore, when the reference image is stored in the memory, except that the storage page of the first column of the first row does not need to repeatedly store data, the storage pages of the other rows all repeatedly store the bottom macroblock data of the previous storage page, And the storage pages of each column repeatedly store the rightmost macroblock data of the previous storage page, as shown in the hatched area in FIG. 4 . Taking the size of the reference image as 720*576 pixels as an example, when the storage method of the present invention is not used, the number of rows of the storage page is 9 rows, and the number of columns is 11.25 columns. After using the storage method of the present invention, since the storage pages of each row and each column must repeatedly store the data of a macroblock, the number of rows and columns of the storage page is increased by about 1/3, that is, the total number of rows is increased to 12 , and the total number of columns increases to 15. Although the capacity of the memory needs to be increased, the boundaries of each memory page can be eliminated. As shown in FIG. 4(B), there is only one storage page SPO defined by the system, so when the system reads any prediction block, there will be no page change action, and the speed of reading the prediction block is improved. In this way, after the reference image is stored in the memory by the method of the present invention, when the system reads any prediction block, there will be no page change action, so in high-bandwidth real-time applications, it will be able to overcome the bandwidth limitation caused by the page change action. bottleneck problem.

Although the present invention has been described above with examples, the scope of the present invention is not limited thereto. Those skilled in the art can make various changes or changes as long as they do not depart from the concept of the present invention.

Claims (20)

1. access method of storage, with the block be unit from the storage access reference picture, comprise the following step:

Deciding step determines the full-size of a prediction block;

Storing step, according to the border of page or leaf, repeated storage segment boundary data; And

Read step if described prediction block is positioned on the page boundary with described repeated storage data, then reads described prediction block from the page or leaf of all data with described prediction block; With the elimination loss of skipping.

2. access method of storage as claimed in claim 1, the full-size of wherein said prediction block is maximum height.

3. access method of storage as claimed in claim 2, wherein said storing step are that repeated storage is less than or equal to described maximum height-1 in the segment boundary data of horizontal boundary and the row of repeated storage.

4. access method of storage as claimed in claim 3, the height of each page on wherein said horizontal repeated storage border is more than or equal to the maximum height of prediction block.

5. access method of storage as claimed in claim 3 in the wherein said read step, when described prediction block can stride across horizontal page boundary, then reads described prediction block from the page or leaf with repeated storage data boundary.

6. access method of storage as claimed in claim 1, the full-size of wherein said prediction block is breadth extreme.

7. access method of storage as claimed in claim 6, wherein said storing step are that repeated storage is less than or equal to described breadth extreme-1 in the segment boundary data of vertical boundary and the row of repeated storage.

8. access method of storage as claimed in claim 7, the width of each page on wherein said vertical repeated storage border is more than or equal to the breadth extreme of prediction block.

9. access method of storage as claimed in claim 8 in the wherein said read step, when described prediction block can stride across vertical page boundary, then reads described prediction block from the page or leaf with repeated storage data boundary.

10. access method of storage as claimed in claim 1, wherein said prediction block full-size comprises maximum height and breadth extreme.

11. access method of storage as claimed in claim 10, wherein said read step are the segment boundary data of repeated storage vertical boundary and horizontal boundary.

12. access method of storage as claimed in claim 11, the height of wherein said page or leaf and width are respectively more than or equal to the maximum height and the breadth extreme of described prediction block.

13. access method of storage as claimed in claim 12 in the wherein said read step, when described prediction block can stride across vertical and/or horizontal memory page border, then reads described prediction block from the page or leaf with repeated storage data boundary.

14. an access method of storage, with the block be unit with reference pictures store in storer, comprise the following step:

Deciding step determines the full-size of a prediction block;

Storing step, according to the border of page or leaf, repeated storage segment boundary data.

15. memory storage methods as claimed in claim 14, wherein said full-size is maximum height.

16. being repeated storage, memory storage methods as claimed in claim 15, wherein said repeated storage step be less than or equal to described maximum height-1 in the segment boundary data of horizontal boundary and the row of repeated storage.

17. memory storage methods as claimed in claim 16, the height of each page on wherein said horizontal repeated storage border is less than or equal to the maximum height of prediction block.

18. memory storage methods as claimed in claim 14, wherein said full-size is breadth extreme.

19. memory storage methods as claimed in claim 18, wherein said repeated storage step are repeated storage at the row of the segment boundary data of vertical boundary and repeated storage more than or equal to described breadth extreme-1.

20. memory storage methods as claimed in claim 19, the width of each page on wherein said vertical repeated storage border is more than or equal to the width of prediction block.

Images